A rice breeding water and fertilizer intelligent regulation and control method and system based on unmanned aerial vehicle remote sensing

Abstract

The present application relates to the field of wisdom agricultural technology, and more particularly to a rice breeding water and fertilizer intelligent regulation and control method and system based on unmanned aerial vehicle remote sensing, comprising: collecting paddy field multi-modal remote sensing data; constructing a spatio-temporal adaptive multi-modal fusion network and a double-branch neural network to predict leaf nitrogen content, soil moisture content and stress level; constructing a water-nitrogen coupling physiological deficit quantification model to generate a physiological deficit index; constructing a digital twin scene and integrating a crop growth model to deduce, taking the physiological deficit index and the stress level as initial states, to obtain simulated leaf nitrogen content, soil moisture content and stress level; generating an optimal decision based on multi-objective optimization; and realizing closed-loop regulation and control through a multi-agent collaborative control system. The present application integrates multi-modal remote sensing data, a water-nitrogen coupling model and digital twin deduction to realize accurate perception, intelligent decision-making and collaborative execution of rice water and nitrogen management, and significantly improves water and fertilizer utilization efficiency and intelligent level.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, specifically to a method and system for intelligent water and fertilizer regulation in rice breeding based on unmanned aerial vehicle (UAV) remote sensing. Background Technology

[0002] Rice is one of my country's major food crops, and water and fertilizer management during its propagation process directly affects seedling quality, seedling establishment rate, and final yield. Traditional water and fertilizer management mainly relies on the experience and judgment of agricultural technicians, which is difficult to cope with the spatial heterogeneity of soil nutrients and the dynamic changes in meteorological conditions, resulting in problems such as low water and nitrogen use efficiency and serious non-point source pollution.

[0003] In recent years, UAV remote sensing technology and machine learning algorithms have been widely used in the field of crop nutrition diagnosis. In existing technologies, there are already schemes that use UAV multispectral images to invert crop nitrogen content or water stress index, generate fertilization or irrigation prescription maps, and guide variable operations. However, these methods have the following shortcomings: (1) Water and fertilizer regulation is usually decided separately, ignoring the water-nitrogen coupling effect and making it difficult to achieve synergistic optimization; (2) Remote sensing diagnostic models are mostly static and fail to be dynamically corrected by combining crop growth mechanism models, and the decision lacks a physiological basis; (3) There is a lack of closed-loop feedback mechanism between the issuance of decision instructions and the execution equipment, and the accuracy of operations is difficult to guarantee; (4) The fusion depth of multi-source data (remote sensing, meteorology, soil) is insufficient, and the ability to represent spatiotemporal heterogeneity is limited.

[0004] To address the aforementioned issues, this invention proposes a method and system for intelligent water and fertilizer regulation in rice breeding based on UAV remote sensing and multi-agent collaboration. By constructing a water-nitrogen coupled physiological deficit quantification model, a multi-agent collaborative decision-making architecture, and an edge-cloud collaborative computing framework, precise water and fertilizer regulation throughout the entire rice breeding cycle can be achieved. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for intelligent regulation of water and fertilizer in rice breeding based on UAV remote sensing, so as to solve the technical problems of water and fertilizer separation decision-making, static diagnostic models, and lack of execution closed loop in the prior art.

[0006] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution: A method for intelligent water and fertilizer regulation in rice breeding based on UAV remote sensing includes the following steps: Step S1: Collect multimodal remote sensing data of paddy fields using an unmanned aerial vehicle platform equipped with a multispectral sensor and a thermal infrared imager, and perform preprocessing; Step S2: Construct a spatiotemporal adaptive multimodal fusion network, extract spectral features and spatial features, fuse meteorological time-series data and soil baseline data to generate a comprehensive feature vector, and construct a dual-branch neural network, wherein the first branch neural network performs a regression task to predict the leaf nitrogen content at the current moment based on the comprehensive feature vector. and soil moisture content The second branch of the neural network is for classification, predicting the current stress level of rice based on the comprehensive feature vector. ; Step S3: Based on the critical nitrogen concentration dilution model and the water deficit index, construct a water-nitrogen coupled physiological deficit quantification model, and use the predictions from step S2. and Calculate nitrogen deficit and water deficit The physiological deficit index at the current moment is generated by correcting the coupling effect matrix. ; Step S4: Construct a digital twin scenario for rice breeding, integrate a crop growth model, and use the physiological deficit index output in step S3. and the stress level output in step S2 As an initial state, multi-strategy simulations are conducted in a digital twin scenario to simulate the crop's response process under different water and nitrogen regulation schemes, and the leaf nitrogen content simulated by the crop growth model during the simulation is obtained. Soil moisture content and the level of coercion ; Step S5: Based on multi-objective optimization, use the derived... , , The goal is to minimize the deviation from the ideal target, thereby generating optimal irrigation and fertilization decisions. Step S6: Deploy a multi-agent collaborative control system, including a decision agent, a perception agent, and an execution agent, to achieve closed-loop control of the issuance, execution, and status feedback of the optimal irrigation and fertilization decisions generated in step S5 through a collaborative mechanism.

[0007] As a preferred embodiment of the present invention, the preprocessing in step S1 includes radiometric calibration, atmospheric correction, geometric correction, and rice canopy extraction based on a semantic segmentation network of deep learning. The semantic segmentation network adopts the U-Net architecture, with multispectral pseudo-color images as input and canopy segmentation masks as output.

[0008] As a preferred embodiment of the present invention, the spatiotemporal adaptive multimodal fusion network in step S2 includes: Spectral feature extraction branch: A one-dimensional convolutional neural network is used to extract the temporal features of multispectral bands and vegetation indices; Spatial feature extraction branch: Residual network is used to extract the spatial distribution features of the canopy; Adaptive Feature Fusion Module: Dynamically fuses spectral features, spatial features, meteorological time-series data, and soil baseline data based on an attention mechanism.

[0009] As a preferred embodiment of the present invention, the critical nitrogen concentration dilution model formula in step S3 is as follows: ,in The critical nitrogen concentration is (g / kg). Aboveground biomass (kg / ha). , Variety-specific parameters; nitrogen deficit That is, the critical nitrogen accumulation rate minus the actual nitrogen accumulation rate, where The leaf nitrogen content predicted in step S2; Water deficit ,in The target soil moisture content (%) is set according to the rice variety and growth stage. This refers to the soil moisture content predicted in step S2.

[0010] As a preferred embodiment of the present invention, the method for constructing the coupling effect matrix in step S3 is as follows: fitting a correction function of water stress on nitrogen absorption based on field test data. and the correction function of nitrogen stress on the transpiration coefficient Using the correction function to and Perform coupling correction to generate a physiological deficit index. ,in Soil moisture content (%) Leaf nitrogen content (g / kg) , , , These are empirical parameters.

[0011] As a preferred embodiment of the present invention, the crop growth model integrated into the digital twin scenario in step S4 is a lightweight version of the WOFOST or DSSAT model, adapted to local varieties and climatic conditions through transfer learning; the multi-objective optimization in step S5 aims to maximize yield, minimize water and nitrogen input, and minimize the inferred stress level, constructing a comprehensive objective function: ; in, , These are the ideal leaf nitrogen content (g / kg) and ideal soil moisture content (%) set according to the target yield. , , These represent the quantified values ​​of leaf nitrogen content, soil moisture content, and stress level output from a crop growth model simulation under a specific water and nitrogen regulation scheme in a digital twin scenario. , , These are the weighting coefficients.

[0012] In a preferred embodiment of the present invention, the first branch neural network of the regression task in step S2 adopts a multi-task learning architecture, and outputs the leaf nitrogen content through two parallel fully connected subnetworks respectively. and soil moisture content Its loss function is the weighted mean square error: ; in , The weights are used as coefficients; the second branch of the neural network for the classification task uses a softmax output layer to output the probability distribution of the stress level, and its loss function is cross-entropy. ; in For the number of coercion level categories, One-hot encoding of the real label. Let c be the probability of the class predicted by the model; the total loss function during training is a weighted sum of the regression loss and the classification loss: ,in These are the weighting coefficients.

[0013] As a preferred embodiment of the present invention, leaf nitrogen content The calculation formula is: ; The soil moisture content The calculation formula is: ; in For the comprehensive feature vector, , For trainable weights and biases, For the Sigmoid function, superscript and These represent the parameters of the first and second layers of the neural network, respectively.

[0014] As a preferred embodiment of the present invention, the present invention provides a rice breeding water and fertilizer intelligent control system based on UAV remote sensing, applied to a rice breeding water and fertilizer intelligent control method based on UAV remote sensing, characterized in that: it includes a multi-agent collaborative control system, the multi-agent collaborative control system comprising: Decision-making intelligent agent: Deployed in the cloud or edge computing nodes, responsible for task planning, decision generation and strategy optimization; Sensing agents: Deployed on drones and field sensor networks to collect crop growth data and environmental data in real time; The executing agent includes a variable irrigation controller and a variable fertilization controller, which are used to receive decision commands and convert them into equipment control parameters; Collaboration mechanism: Communication between agents is realized based on the publish-subscribe model. The executing agent provides real-time feedback on the job status, and the sensing agent uploads monitoring data to trigger decision updates. The sensing agent includes a drone remote sensing module, which is equipped with a multispectral sensor and a thermal infrared imager for collecting rice paddy image data. The decision-making agent includes edge computing nodes and a cloud management platform. The edge computing nodes are deployed in the field and have a built-in lightweight dual-branch neural network model for real-time output of leaf nitrogen content. Soil moisture content and coercion level The cloud management platform integrates digital twin scenarios, crop growth models, and a multi-objective optimization engine for use in... , , The system simulates the initial state and generates optimal irrigation and fertilization decisions. The executing agent also includes a variable irrigation execution module and a variable fertilization execution module. The variable irrigation execution module includes an irrigation network, a solenoid valve group, and a frequency converter, and is used to execute variable irrigation according to the decision. The variable fertilization execution module includes a fertilizer injection system, an EC / pH sensor, and a frequency converter, and is used to execute variable fertilization according to the decision. The multi-agent collaborative control system also includes a multi-agent collaborative control bus, which connects the decision-making agent, the perception agent, and the execution agent to realize data interaction and command issuance.

[0015] As a preferred embodiment of the present invention, the dual-branch neural network model built into the edge computing node supports online incremental updates.

[0016] Compared with the prior art, the present invention has the following advantages: This invention constructs a water-nitrogen coupling effect matrix to quantify the water-nitrogen interaction mechanism, avoids decision-making bias caused by single-factor regulation, achieves synergistic optimization of water and fertilizer resources, and integrates the critical nitrogen concentration dilution model with the crop water stress index to transform remote sensing inversion results into physiologically significant water and nitrogen deficits, making the decision-making basis more scientific. Attached Figure Description

[0017] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for intelligent regulation of water and nitrogen in rice breeding based on UAV remote sensing, as disclosed in an embodiment of this application. Figure 2 This is a schematic diagram of the spatiotemporal adaptive multimodal fusion network structure disclosed in the embodiments of this application; Figure 3 This is a schematic diagram of the water-nitrogen coupling physiological deficit quantification model disclosed in an embodiment of this application; Figure 4 This is a schematic diagram of a smart water and nitrogen regulation system for rice breeding based on UAV remote sensing, as disclosed in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, this invention provides a method for intelligent water and fertilizer regulation in rice breeding based on UAV remote sensing, comprising the following steps: Step S1: Collect multimodal remote sensing data of paddy fields using an unmanned aerial vehicle platform equipped with a multispectral sensor and a thermal infrared imager, and perform preprocessing; The preprocessing in step S1 includes radiometric calibration, atmospheric correction, geometric correction, and rice canopy extraction based on a deep learning-based semantic segmentation network. The semantic segmentation network adopts the U-Net architecture, with multispectral pseudo-color images as input and canopy segmentation masks as output.

[0021] Low-altitude remote sensing operations were conducted on rice paddies using a drone platform (such as the DJI M300 RTK) equipped with a multispectral sensor (such as the RedEdge-MX, with bands including blue (475nm), green (560nm), red (668nm), red-edge (717nm), and near-infrared (842nm)) and a thermal infrared imager (such as the FLIR Vue Pro R, with a resolution of 640×512 and thermal sensitivity <50mK). The data collected included multispectral and thermal infrared images across multiple bands. The raw remote sensing data underwent preprocessing, including radiometric calibration (converting DN values ​​to radiance), atmospheric correction (e.g., using the MODTRAN model and the FLAASH module in ENVI software), and geometric correction (based on drone POS data and ground control points, generating orthorectified images using Pix4Dmapper software) to eliminate sensor errors, atmospheric scattering, and absorption effects, thereby obtaining accurate geospatial information.

[0022] Furthermore, to accurately obtain information about the rice canopy, it is necessary to separate the canopy from the background (such as water and soil). This embodiment employs a deep learning-based semantic segmentation network, specifically the U-Net architecture. It takes a multispectral pseudo-color image (a combination of near-red, red, and green bands) as input and outputs a canopy segmentation mask of the same size as the input image. In the mask, a pixel value of 1 represents the rice canopy, and 0 represents the background. The U-Net network structure includes an encoder (4 downsampling operations, each using two 3×3 convolutions + ReLU + 2×2 max pooling, with the number of channels increasing layer by layer from 64 to 512) and a decoder (4 upsampling operations, each using a 2×2 transposed convolution + skip connections + two 3×3 convolutions + ReLU). Finally, the segmentation result is output through a 1×1 convolution and sigmoid activation. This mask allows for the extraction of a clean rice canopy region for subsequent spectral feature extraction.

[0023] Step S2: Construct a spatiotemporal adaptive multimodal fusion network to extract spectral and spatial features, fuse meteorological time-series data and soil baseline data, generate a comprehensive feature vector, and construct a dual-branch neural network. The first branch of the neural network performs a regression task, predicting the leaf nitrogen content at the current moment based on the comprehensive feature vector. and soil moisture content The second branch of the neural network is for classification, predicting the current stress level of rice based on the comprehensive feature vector. ; The spatiotemporal adaptive multimodal fusion network in step S2 includes: Spectral feature extraction branch: A one-dimensional convolutional neural network is used to extract the temporal features of multispectral bands and vegetation indices; Spatial feature extraction branch: Residual network is used to extract the spatial distribution features of the canopy; Adaptive Feature Fusion Module: Dynamically fuses spectral features, spatial features, meteorological time-series data, and soil baseline data based on an attention mechanism.

[0024] This invention constructs a spatiotemporal adaptive multimodal fusion network (e.g. Figure 2 This method extracts comprehensive features from preprocessed multimodal data and predicts current leaf nitrogen content (LNC), soil water content (SWC), and stress level in real time. The specific structure and training details of the network are as follows: Spectral Feature Extraction Branch: Based on the canopy region extracted in step S1, multiple vegetation indices (such as NDVI, NDRE, GNDVI, EVI, etc.) are calculated. The spectral reflectance values ​​of multiple bands and the calculated vegetation indices are combined to form a time series (for historical remote sensing data of the same field, with a time window length of 5 days) or a one-dimensional vector (for a single remote sensing event), which is then input into a one-dimensional convolutional neural network (1D-CNN). This 1D-CNN contains three convolutional layers, each with a kernel size of 3, a stride of 1, and 32, 64, and 128 channels respectively. Each layer is followed by a ReLU activation function and a max-pooling layer (pooling size 2). Finally, a 256-dimensional spectral feature vector is obtained through a global average pooling layer.

[0025] Spatial Feature Extraction Branch: Using five bands of the multispectral image as five-channel input, a ResNet-18 residual network is employed to extract the spatial texture and distribution features of the rice canopy. Specifically, the first four residual blocks of ResNet-18 are used, outputting a feature map with a size 1 / 16 of the original image and 512 channels. Global average pooling is then applied to obtain a 512-dimensional spatial feature vector. These features reflect the spatial heterogeneity of crop growth.

[0026] The adaptive feature fusion module concatenates the extracted spectral feature vector (256-dimensional) and spatial feature vector (512-dimensional) with synchronously collected meteorological time-series data (temperature, humidity, radiation, and rainfall, normalized to form a 4-dimensional vector) and basic soil data (such as soil type, bulk density, and initial organic matter content, one-hot encoded to form a 10-dimensional vector) to obtain an original feature vector of dimension 782. Subsequently, an adaptive fusion module based on an attention mechanism is used. This module contains a two-layer fully connected network (256 neurons + ReLU in the first layer, and 782 neurons + Sigmoid in the second layer) to dynamically generate weights for each dimension of the original feature vector. Multiplying these weights by the original features yields a weighted comprehensive feature vector F (still 782-dimensional). The attention weights are automatically learned through end-to-end training to highlight features that contribute significantly to the prediction task.

[0027] Dual-branch neural network: Construct a dual-branch neural network with a shared bottom layer (one fully connected layer, 256 neurons + ReLU). The shared bottom layer extracts general features from the comprehensive feature vector F, and then inputs them into two task-specific branches respectively.

[0028] The first branch (regression task): employs a multi-task learning architecture, comprising two parallel fully connected subnetworks. Shared features are input separately: LNC subnet: Two fully connected layers, the first layer has 128 neurons + ReLU, and the second layer has 1 neuron (linear output), outputting the leaf nitrogen content (LNC). The calculation formula can be expressed as: .

[0029] SWC subnetwork: Two fully connected layers. The first layer has 128 neurons + ReLU, and the second layer has 1 neuron + Sigmoid activation. It outputs soil moisture content (SWC) in the range of 0-1, multiplied by 100% to get the percentage. Calculation formula: .

[0030] The feature vector input to this subnet is typically a composite feature vector extracted and passed from the upstream network (such as a shared bottom layer), with dimensions of [missing information]. (e.g., 256 dimensions).

[0031] superscript and : These represent the parameters of the first layer (hidden layer) and the second layer (output layer), respectively.

[0032] Weight matrix. For example... This is the weight matrix of the first hidden layer, and its shape is... (128 neurons, each neuron connected to the input) (each dimension is connected) It is the weight matrix of the output layer, with the shape of... .

[0033] Bias vector. For example... It is the offset of the first layer, with a length of 128; It is the bias of the output layer, which is a scalar.

[0034] ReLU: The activation function of a linear rectifier unit, defined as follows: , used to introduce nonlinearity.

[0035] The Sigmoid activation function is defined as follows: Compress the output to An interval is suitable for probability or normalized values.

[0036] The first linear transformation yields 128-dimensional intermediate values.

[0037] Apply ReLU activation element by element to the above intermediate values ​​to obtain the hidden layer output (128 dimensions).

[0038] The second linear transformation maps the 128-dimensional vector to a scalar (linear output), which is the predicted value of leaf nitrogen content (unit: g / kg).

[0039] The first two steps are the same as the LNC subnet: the first layer linear transformation + ReLU to obtain the hidden layer output (128 dimensions), and then the second layer linear transformation to obtain a scalar.

[0040] Applying Sigmoid activation to this scalar compresses it to... The interval represents the normalized soil moisture content (0~1).

[0041] Multiply the normalized value by 100 to convert it into soil moisture content as a percentage (0~100%).

[0042] For the comprehensive feature vector, , For trainable weights and biases, This is the Sigmoid function.

[0043] The second branch (classification task): used to predict the stress level of rice (e.g., normal, mild drought, severe drought, nitrogen deficiency, water-nitrogen coupled stress, etc., a total of C=5 categories). This branch consists of two fully connected layers: the first layer has 128 neurons + ReLU, and the second layer has C neurons + softmax, outputting the probability distribution of each category. The stress level S_real is the category with the highest probability.

[0044] The network is trained using a joint loss function. A dataset is constructed using historical field observation data (including synchronized remote sensing data, meteorological and soil data, and measured LNC, SWC, and manually labeled stress levels), and randomly divided into training, validation, and test sets at 80%, 10%, and 10% respectively. During training, the loss function for the regression task is... Weighted mean square error (MSE) of LNC and SWC predictions: Where λ1=0.6, λ2=0.4. Loss for the classification task. Cross-entropy loss: The total loss function is λ3=0.5. The Adam optimizer is used, with an initial learning rate of 0.001, which decays to 0.1 every 20 epochs. The batch size is set to 32, and training is performed for 100 epochs. The loss is monitored on the validation set, with an early stopping time of 10 epochs. After training, the model parameters with the minimum loss on the validation set are saved.

[0045] Specifically, in step S2, the first branch neural network of the regression task adopts a multi-task learning architecture, outputting the leaf nitrogen content through two parallel fully connected subnetworks. and soil moisture content Its loss function is the weighted mean square error: ; in , The weights are used as coefficients; the second branch of the neural network for the classification task uses a softmax output layer to output the probability distribution of the stress level, and its loss function is cross-entropy. ; in For the number of coercion level categories, One-hot encoding of the real label. Let c be the probability of the class predicted by the model; the total loss function during training is a weighted sum of the regression loss and the classification loss: ,in These are the weighting coefficients.

[0046] Step S3: Based on the critical nitrogen concentration dilution model and the water deficit index, construct a water-nitrogen coupled physiological deficit quantification model, and use the predictions from step S2. and Calculate nitrogen deficit and water deficit The physiological deficit index at the current moment is generated by correcting the coupling effect matrix. ; The critical nitrogen concentration dilution model formula in step S3 is: ,in The critical nitrogen concentration is (g / kg). Aboveground biomass (kg / ha). , Variety-specific parameters; nitrogen deficit That is, the critical nitrogen accumulation rate minus the actual nitrogen accumulation rate, where The leaf nitrogen content predicted in step S2; Water deficit ,in The target soil moisture content (%) is set according to the rice variety and growth stage. This refers to the soil moisture content predicted in step S2.

[0047] The method for constructing the coupling effect matrix in step S3 is as follows: A correction function for the effect of water stress on nitrogen uptake is fitted based on field trial data. and the correction function of nitrogen stress on the transpiration coefficient Using the correction function to and Perform coupling correction to generate a physiological deficit index. ,in Soil moisture content (%) Leaf nitrogen content (g / kg) , , , As empirical parameters, the above model parameters (a, b, k, SWC0, α, β) are all obtained by fitting multi-year field trial data. The trial data covers the main local rice varieties and typical climatic conditions to ensure that the model has a physiological basis and regional adaptability.

[0048] This invention constructs a physiological deficit index P that can reflect the interaction between water and nitrogen (see...). Figure 3 First, based on the LNC predicted in step S2, the nitrogen deficit ΔN is calculated. This requires the use of a critical nitrogen concentration dilution model, which describes the minimum nitrogen concentration required for the crop to reach its maximum growth rate at a given biomass. The model formula is: ,in The critical nitrogen concentration (g / kg) The aboveground biomass (kg / ha, obtained by establishing a regression model between vegetation indices (such as NDVI) obtained through remote sensing inversion and measured biomass) is given. a and b are empirical parameters dependent on rice varieties (e.g., for indica rice, a=3.61, b=0.36; for japonica rice, a=3.25, b=0.32, obtained by fitting multi-year field trial data). The nitrogen deficit ΔN is the difference between the critical nitrogen accumulation and the actual nitrogen accumulation. .

[0049] At the same time, the water deficit ΔW is calculated. Where SWC is the soil moisture content predicted in step S2. It is the target soil moisture content (%) set according to different growth stages of rice (such as tillering stage, jointing stage, heading stage, and grain filling stage). The target value is provided by the agronomic knowledge base (e.g., 80%±5% for tillering stage and 85%±5% for jointing stage).

[0050] However, water and nitrogen deficits are mutually influential. Insufficient water limits nitrogen uptake by roots, while nitrogen deficiency weakens root growth and leaf photosynthesis, thus affecting crop water use efficiency. Therefore, this step introduces a coupling effect matrix to correct for ΔN and ΔW. This matrix, based on multi-year, multi-location field experimental data, includes two correction functions: Correction function for nitrogen uptake under water stress: The sigmoid function represents the condition when the soil moisture content (SWC) is below a threshold. At this point, nitrogen absorption efficiency begins to decrease. (Parameter k=0.5) =60% was obtained by fitting experimental data.

[0051] Nitrogen stress correction function for transpiration coefficient: This function indicates that when the leaf nitrogen content (LNC) decreases, the crop's transpiration (i.e., water requirement) decreases accordingly. The parameters α=0.8 and β=0.3 were obtained by fitting experimental data.

[0052] Finally, the physiological deficit index P at the current moment is generated, which is a comprehensive scalar: P = ω1·(ΔN·f(SWC)) + ω2·(ΔW·g(LNC)), where ω1=0.5 and ω2=0.5 are weighting coefficients. The larger the P value, the more severe the current comprehensive physiological deficit of the crop.

[0053] Step S4: Construct a digital twin scenario for rice breeding, integrate a crop growth model, and use the physiological deficit index output in step S3. and the stress level output in step S2 As an initial state, multi-strategy simulations are conducted in a digital twin scenario to simulate the crop's response process under different water and nitrogen regulation schemes, and the leaf nitrogen content simulated by the crop growth model during the simulation is obtained. Soil moisture content and the level of coercion ; This invention constructs a digital twin scene corresponding to a real rice paddy. The core of this scene is the integration of a lightweight crop growth model (such as a simplified version of WOFOST). This model, through transfer learning, utilizes local meteorological and field management data from the past five years for parameter calibration to adapt to specific local rice varieties and climatic conditions. The WOFOST model primarily simulates processes such as crop canopy development, photosynthesis, dry matter accumulation and distribution, soil water balance, and nitrogen cycling. The digital twin scene is developed using the Unity3D engine, providing a 3D representation of the rice paddy plots, crop growth, and environmental elements.

[0054] The initial state of the simulation is the stress level output in step S2. And the physiological deficit index P output from step S3 is set. Specifically, P is converted into initial parameters for the crop growth model (such as initial leaf area index, soil moisture content, and plant nitrogen content). Then, in the digital twin scenario, multiple different water and nitrogen control schemes are set for the next 7 days (e.g., irrigation volume of 50-200 m³). 3 Within a range of / ha, 10m 3 Fertilizer application rates were set at 10 kg / ha intervals within the range of 50-200 kg / ha, resulting in 256 possible schemes. For each scheme, a crop growth model was used to simulate the crop growth process under that scheme, and changes in key state variables were recorded, especially the simulated leaf nitrogen content. Soil moisture content and the stress level calculated within the crop model (Model output value).

[0055] The integrated WOFOST crop growth model calculates water stress factors daily. (Actual transpiration / potential transpiration) and nitrogen stress factors (Actual leaf nitrogen content / critical nitrogen content). This example defines the comprehensive stress factor. Coercion Level Quantification Value Through the A linear transformation yields: The value ranges from 0 to 100, with a larger value indicating more severe stress. This value is directly used as one of the inputs for multi-objective optimization in step S5.

[0056] Step S5: Based on multi-objective optimization, use the derived... , , The goal is to minimize the deviation from the ideal target, thereby generating optimal irrigation and fertilization decisions. The crop growth model integrated into the digital twin scenario in step S4 is a lightweight version of the WOFOST or DSSAT model, adapted to local varieties and climatic conditions through transfer learning; the multi-objective optimization in step S5 aims to maximize yield, minimize water and nitrogen input, and minimize the inferred stress level, constructing a comprehensive objective function: ; in, , These are the ideal leaf nitrogen content (g / kg) and ideal soil moisture content (%) set according to the target yield. , , These represent the quantified values ​​of leaf nitrogen content, soil moisture content, and stress level output from a crop growth model simulation under a specific water and nitrogen regulation scheme in a digital twin scenario. , , These are the weighting coefficients.

[0057] The derivation results in step S4 are used to find the optimal water and nitrogen regulation scheme. This is a multi-objective optimization problem, aiming to balance conflicting objectives, such as: Yield maximization (reflected in the growth model, represented by the final biomass or yield simulation value).

[0058] Minimize water and nitrogen input (minimize irrigation water and fertilizer application).

[0059] Optimization of crop health status (stress levels in the extrapolation process) (As low as possible).

[0060] To transform a multi-objective problem into a single-objective problem for solution, this invention constructs a comprehensive objective function and minimizes the function. : in, and These are the ideal average leaf nitrogen content and ideal average soil moisture content, respectively, set according to the target yield (e.g., when the target yield is 10 t / ha). =25g / kg, =80%) , , These represent the average values ​​of the model output over the projected period under a specific control scheme. , , Therefore, we get α, β, and γ are weighting coefficients (α=0.3, β=0.3, γ=0.4). The Non-Dominated Sorting Genetic Algorithm (NSGA-II) is used for optimization, with a population size of 100, 50 iterations, a crossover probability of 0.8, and a mutation probability of 0.1. The algorithm searches the solution space for the solution that minimizes F; the irrigation and fertilization amounts corresponding to this solution are the optimal decisions for this decision cycle.

[0061] Step S6: Deploy a multi-agent collaborative control system, including a decision agent, a perception agent, and an execution agent, to achieve closed-loop control of the issuance, execution, and status feedback of the optimal irrigation and fertilization decisions generated in step S5 through a collaborative mechanism.

[0062] Finally, the optimal irrigation and fertilization decisions generated in step S5 are sent to the field execution equipment. To achieve efficient and reliable closed-loop control, this step deploys a multi-agent cooperative control system (such as...). Figure 4 As shown), the system includes: Perception agents: Deployed on drones and field sensor networks (such as soil temperature and humidity sensors and small weather stations), they are responsible for collecting crop growth data and environmental data in real time and publishing them to the collaborative bus via the MQTT protocol.

[0063] Decision-making agent: Edge computing nodes deployed in the cloud or in the field (such as NVIDIA Jetson Xavier NX). These edge computing nodes have a built-in lightweight dual-branch neural network model (i.e., the model trained in step S2), capable of rapidly inferring from real-time sensor data and outputting the current LNC, SWC, and... The model supports online incremental updates. When newly collected measured data (such as plant sampling and testing results) accumulates to a certain amount (e.g., 100 sets), model fine-tuning is triggered. The model is updated for several epochs using the new data to continuously improve its accuracy. The cloud management platform (deployed on Alibaba Cloud ECS) is responsible for running more complex digital twin models and multi-objective optimization engines, generating optimal decisions, and issuing decision instructions.

[0064] The execution agent includes a variable irrigation controller and a variable fertilization controller. The variable irrigation execution module consists of a zone-controlled irrigation network, a fast-response solenoid valve assembly, and a variable frequency controller that can adjust irrigation pressure. The variable fertilization execution module, based on the irrigation system, integrates a fertilization system controlled by a variable frequency pump, flow meter, and EC / pH sensor. It can precisely adjust the concentration and flow rate of fertilizer stock solution injected into the irrigation water according to decision commands, achieving simultaneous and precise application of water and fertilizer.

[0065] Collaboration Mechanism: Communication between all agents is achieved using a publish-subscribe model, with JSON as the message format. For example, the perception agent publishes real-time data to the topic `sensor_data`. The decision agent on the edge computing node subscribes to this topic, infers the LNC and SWC, and then publishes them to the topic `real_time_status`. The cloud platform decision agent subscribes to this topic, combines it with other data, and after deduction and optimization, publishes the optimal decision to the topic `optimal_decision`. The execution agent subscribes to this topic, receives instructions, starts the operation, and publishes the operation status (such as valve opening and instantaneous flow) to the topic `actuator_status` in real time, achieving full transparency and closed-loop feedback.

[0066] The method provided in this embodiment constructs a closed loop for intelligent regulation of water and nitrogen in rice through the above six steps, from data perception, model diagnosis, quantitative evaluation, virtual simulation, optimization decision-making to intelligent execution, which significantly improves the scientific nature and accuracy of water and fertilizer management.

[0067] like Figure 4 As shown, this invention provides an intelligent water and fertilizer control system for rice breeding based on UAV remote sensing, applied to an intelligent water and fertilizer control method for rice breeding based on UAV remote sensing. Its characteristic is that it includes a multi-agent collaborative control system, which comprises: Decision-making intelligent agent: Deployed in the cloud or edge computing nodes, responsible for task planning, decision generation and strategy optimization; Sensing agents: Deployed on drones and field sensor networks to collect crop growth data and environmental data in real time; The executing agent includes a variable irrigation controller and a variable fertilization controller, which are used to receive decision commands and convert them into equipment control parameters; Collaboration mechanism: Communication between agents is realized based on the publish-subscribe model, the executing agent provides real-time feedback on the operation status, and the sensing agent uploads monitoring data to trigger decision updates; The sensing agent includes a drone remote sensing module, which is equipped with a multispectral sensor and a thermal infrared imager to collect rice paddy image data; The decision-making agent consists of edge computing nodes and a cloud management platform. The edge computing nodes are deployed in the field and have a built-in lightweight dual-branch neural network model to output the leaf nitrogen content at the current moment in real time. Soil moisture content and coercion level The cloud management platform integrates digital twin scenarios, crop growth models, and a multi-objective optimization engine for use in... , , The system simulates the initial state and generates optimal irrigation and fertilization decisions. The executing agent also includes a variable irrigation execution module and a variable fertilization execution module. The variable irrigation execution module includes an irrigation network, a solenoid valve group, and a variable frequency controller, which is used to execute variable irrigation according to decisions. The variable fertilization execution module includes a fertilizer injection system, an EC / pH sensor, and a variable frequency pump group, which is used to execute variable fertilization according to decisions. The multi-agent cooperative control system also includes a multi-agent cooperative control bus, which connects the decision-making agent, the perception agent, and the execution agent to realize data interaction and command issuance.

[0068] The dual-branch neural network model built into the edge computing node supports online incremental updates.

[0069] The sensing agent comprises a drone remote sensing module and a fixed field sensor network. The drone remote sensing module, equipped with a multispectral sensor and a thermal infrared imager, is used to periodically or irregularly collect high-resolution image data of the paddy field. The field sensor network includes soil temperature and humidity sensors, conductivity sensors, and farmland micro-weather stations buried in the root zone for continuous monitoring of the soil environment and field microclimate. All sensing data is collected via 4G / 5G wireless transmission.

[0070] The decision-making agent consists of edge computing nodes and a cloud management platform. Edge computing nodes are deployed in fields (e.g., in intelligent control cabinets on field ridges), using the NVIDIA Jetson Xavier NX platform. They incorporate a pre-trained lightweight dual-branch neural network model (specific structure as shown in step S2 of Example 1), enabling rapid inference from real-time sensor data and outputting the current LNC, SWC, and S_real values ​​in real time. This model supports online incremental updates, continuously optimizing its accuracy based on newly collected measured data (e.g., plant sampling and testing results). The cloud management platform (deployed on Alibaba Cloud ECS, configured with 8 CPU cores / 32GB memory) boasts enhanced computing and storage capabilities, integrating a digital twin scenario, a complete crop growth model (e.g., WOFOST), and a multi-objective optimization engine (implemented by NSGA-II). Starting with the real-time state output by the edge computing nodes, it extrapolates future scenarios and performs global optimization to generate optimal irrigation and fertilization decisions.

[0071] The execution agent comprises a variable irrigation execution module and a variable fertilization execution module. The variable irrigation execution module consists of a zone-controlled irrigation network, a fast-response solenoid valve assembly (each solenoid valve corresponds to a 10m × 10m irrigation zone), and a variable frequency controller (7.5kW) for adjustable irrigation pressure. The variable fertilization execution module, based on the irrigation system, integrates a fertilization system controlled by a variable frequency pump, flow meter, and EC / pH sensor. It can precisely adjust the concentration and flow rate of the fertilizer stock solution injected into the irrigation water according to decision commands, achieving synchronized and precise water and fertilizer application.

[0072] Collaboration Mechanism and Bus: The system constructs a multi-agent collaborative control bus (based on MQTT Broker, deployed on a cloud server) to connect all agents. All agents follow a unified publish-subscribe communication protocol. For example, the perception agent publishes data to the bus, the decision-making agent subscribes to the required data from the bus, processes it, publishes the results, and the execution agent subscribes to decision instructions and reports the job status. This decentralized collaboration mechanism ensures the system's flexibility, scalability, and robustness.

[0073] The system provided in this embodiment perfectly supports the implementation of the method in Embodiment 1 through clear division of labor and efficient collaboration among intelligent agents, realizing a truly intelligent closed loop for water and nitrogen management in rice breeding.

[0074] This invention constructs a water-nitrogen coupling effect matrix to quantify the water-nitrogen interaction mechanism, avoids decision-making bias caused by single-factor regulation, achieves synergistic optimization of water and fertilizer resources, and integrates the critical nitrogen concentration dilution model with the crop water stress index to transform remote sensing inversion results into physiologically significant water and nitrogen deficits, making the decision-making basis more scientific.

[0075] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A method for intelligent water and fertilizer regulation in rice breeding based on UAV remote sensing, characterized in that, Includes the following steps: Step S1: Collect multimodal remote sensing data of paddy fields using an unmanned aerial vehicle platform equipped with a multispectral sensor and a thermal infrared imager, and perform preprocessing; Step S2: Construct a spatiotemporal adaptive multimodal fusion network, extract spectral features and spatial features, fuse meteorological time-series data and soil baseline data to generate a comprehensive feature vector, and construct a dual-branch neural network, wherein the first branch neural network performs a regression task to predict the leaf nitrogen content at the current moment based on the comprehensive feature vector. and soil moisture content The second branch of the neural network is for classification, predicting the current stress level of rice based on the comprehensive feature vector. ; Step S3: Based on the critical nitrogen concentration dilution model and the water deficit index, construct a water-nitrogen coupled physiological deficit quantification model, and use the predictions from step S2. and Calculate nitrogen deficit and water deficit The physiological deficit index at the current moment is generated by correcting the coupling effect matrix. ; Step S4: Construct a digital twin scenario for rice breeding, integrate a crop growth model, and use the physiological deficit index output in step S3. and the stress level output in step S2 As an initial state, multi-strategy simulations are conducted in a digital twin scenario to simulate the crop's response process under different water and nitrogen regulation schemes, and the leaf nitrogen content simulated by the crop growth model during the simulation is obtained. Soil moisture content and the level of coercion ; Step S5: Based on multi-objective optimization, use the derived... , , The goal is to minimize the deviation from the ideal target, thereby generating optimal irrigation and fertilization decisions. Step S6: Deploy a multi-agent collaborative control system, including a decision agent, a perception agent, and an execution agent, to achieve closed-loop control of the issuance, execution, and status feedback of the optimal irrigation and fertilization decisions generated in step S5 through a collaborative mechanism.

2. The intelligent water and fertilizer regulation method for rice breeding based on UAV remote sensing according to claim 1, characterized in that: The preprocessing in step S1 includes radiometric calibration, atmospheric correction, geometric correction, and rice canopy extraction based on a deep learning-based semantic segmentation network. The semantic segmentation network adopts the U-Net architecture, with multispectral pseudo-color images as input and canopy segmentation masks as output.

3. The intelligent water and fertilizer regulation method for rice breeding based on UAV remote sensing according to claim 2, characterized in that: The spatiotemporal adaptive multimodal fusion network in step S2 includes: Spectral feature extraction branch: A one-dimensional convolutional neural network is used to extract the temporal features of multispectral bands and vegetation indices; Spatial feature extraction branch: Residual network is used to extract the spatial distribution features of the canopy; Adaptive Feature Fusion Module: Dynamically fuses spectral features, spatial features, meteorological time-series data, and soil baseline data based on an attention mechanism.

4. The intelligent water and fertilizer regulation method for rice breeding based on UAV remote sensing according to claim 3, characterized in that: The critical nitrogen concentration dilution model formula in step S3 is as follows: ,in The critical nitrogen concentration is (g / kg). Aboveground biomass (kg / ha). , Variety-specific parameters; nitrogen deficit That is, the critical nitrogen accumulation rate minus the actual nitrogen accumulation rate, where The leaf nitrogen content predicted in step S2; Water deficit ,in The target soil moisture content (%) is set according to the rice variety and growth stage. This refers to the soil moisture content predicted in step S2.

5. The intelligent water and fertilizer regulation method for rice breeding based on UAV remote sensing according to claim 4, characterized in that: The method for constructing the coupling effect matrix in step S3 is as follows: fitting a correction function of water stress on nitrogen uptake based on field test data. and the correction function of nitrogen stress on the transpiration coefficient Using the correction function to and Perform coupling correction to generate a physiological deficit index. ,in Soil moisture content (%) Leaf nitrogen content (g / kg) , , , These are empirical parameters.

6. The intelligent water and fertilizer regulation method for rice breeding based on UAV remote sensing according to claim 5, characterized in that: The crop growth model integrated into the digital twin scenario in step S4 is a lightweight version of the WOFOST or DSSAT model, adapted to local varieties and climatic conditions through transfer learning; the multi-objective optimization in step S5 aims to maximize yield, minimize water and nitrogen input, and minimize the inferred stress level, constructing a comprehensive objective function: ； in, , These are the ideal leaf nitrogen content (g / kg) and ideal soil moisture content (%) set according to the target yield. , , These represent the quantified values ​​of leaf nitrogen content, soil moisture content, and stress level output from a crop growth model simulation under a specific water and nitrogen regulation scheme in a digital twin scenario. , , These are the weighting coefficients.

7. The intelligent water and fertilizer regulation method for rice breeding based on UAV remote sensing according to claim 6, characterized in that: In step S2, the first branch neural network of the regression task adopts a multi-task learning architecture, outputting the leaf nitrogen content through two parallel fully connected subnetworks. and soil moisture content Its loss function is the weighted mean square error: ； in , The weights are used as coefficients; the second branch of the neural network for the classification task uses a softmax output layer to output the probability distribution of the stress level, and its loss function is cross-entropy. ； in For the number of coercion level categories, One-hot encoding of the real label. Let c be the probability of the class predicted by the model; the total loss function during training is a weighted sum of the regression loss and the classification loss: ,in These are the weighting coefficients.

8. The intelligent water and fertilizer regulation method for rice breeding based on UAV remote sensing according to claim 7, characterized in that: Leaf nitrogen content The calculation formula is: ； The soil moisture content The calculation formula is: ； in For the comprehensive feature vector, , For trainable weights and biases, For the Sigmoid function, superscript and These represent the parameters of the first and second layers of the neural network, respectively.

9. A water and fertilizer intelligent control system for rice breeding based on UAV remote sensing, characterized in that, The method for intelligent water and fertilizer regulation in rice breeding based on UAV remote sensing, as described in any one of claims 1-8, is characterized by: comprising a multi-agent collaborative control system, wherein the multi-agent collaborative control system includes: Decision-making intelligent agent: Deployed in the cloud or edge computing nodes, responsible for task planning, decision generation and strategy optimization; Sensing agents: Deployed on drones and field sensor networks to collect crop growth data and environmental data in real time; The executing agent includes a variable irrigation controller and a variable fertilization controller, which are used to receive decision commands and convert them into equipment control parameters; Collaboration mechanism: Communication between agents is realized based on the publish-subscribe model. The executing agent provides real-time feedback on the job status, and the sensing agent uploads monitoring data to trigger decision updates. The sensing agent includes a drone remote sensing module, which is equipped with a multispectral sensor and a thermal infrared imager for collecting rice paddy image data. The decision-making agent includes edge computing nodes and a cloud management platform. The edge computing nodes are deployed in the field and have a built-in lightweight dual-branch neural network model for real-time output of leaf nitrogen content. Soil moisture content and coercion level The cloud management platform integrates digital twin scenarios, crop growth models, and a multi-objective optimization engine for use in... , , The system simulates the initial state and generates optimal irrigation and fertilization decisions. The executing agent also includes a variable irrigation execution module and a variable fertilization execution module. The variable irrigation execution module includes an irrigation network, a solenoid valve group, and a frequency converter, and is used to execute variable irrigation according to the decision. The variable fertilization execution module includes a fertilizer injection system, an EC / pH sensor, and a frequency converter, and is used to execute variable fertilization according to the decision. The multi-agent collaborative control system also includes a multi-agent collaborative control bus, which connects the decision-making agent, the perception agent, and the execution agent to realize data interaction and command issuance.

10. The system according to claim 9, characterized in that, The edge computing node's built-in dual-branch neural network model supports online incremental updates.

Images